Opposition-Based Multi-objective Binary Differential Evolution for Multi-label Feature Selection

摘要

Multi-label learning problem is a data analytic task in which every sample is associated with more than single label. The complexity of such problems declares the importance of feature selection task as a preprocessing step prior for multi-label learning. Feature selection can make a better learning performance both in terms of reducing computational complexity and increasing classification accuracy. Selecting the best subset of features with two objectives, the smaller number of features and higher accuracy of classification can be treated as a binary multi-objective optimization problem. Since feature selection is inherently a binary optimization problem, applying continuous metaheuris-tic algorithms to solve this problem decreases the diversity of solutions in the optimal Pareto-front, because of many-to-one mapping and low exploration power, accordingly. This paper proposed a binary version of Generalized Differential Evolution (BGDE3) for multi-label feature selection based on majority voting of solutions and opposition-based learning (OBL). Experimental results show that the proposed algorithm outperforms the continuous GDE3 for multi-label feature selection.